Spatially Adaptive Classification and Active Learning of Multispectral Data with Gaussian Processes

Multispectral remote sensing images are widely used for automated land use and land cover classification tasks. Remotely sensed images usually cover large geographical areas, and spectral characteristics of each class often varies over time and space. We apply a spatially adaptive classification scheme that models spatial variation with Gaussian processes, and apply uncertainty sampling based active learning algorithm to achieve better classification accuracies with a fewer number of samples. The spatially adaptive classifier shows better performances than the conventional maximum likelihood classifier in both passive and active learning settings, and the active learners achieves better classification accuracies than passive learners with fewer number of samples for both classification algorithms.